The Role of Artificial Intelligence in the Diagnosis, Segmentation, and Prediction of Retinal Vein Occlusion: A...

Retinal vein occlusion (RVO) is the second most common cause of vision loss after diabetic retinopathy. It results from the occlusion of either the central retinal vein or one of its branches. Artificial intelligence (AI), particularly deep learning (DL), has shown great potential in ophthalmology for disease assessment. This review examined how AI has been applied to the diagnosis, segmentation, and treatment prediction of RVO across different imaging modalities. A comprehensive search of PubMed, Scopus, and Google Scholar up to June 19, 2024, identified 2,925 records, of which 23 met the inclusion criteria. Most studies (91%) were published after 2020, reflecting the rapid growth of AI in this field. DL algorithms were used in 87% of studies, mainly convolutional neural networks such as Residual Network, Densely Connected Convolutional Network, and Visual Geometry Group Network. Classification was the most frequent task (78%), followed by segmentation (26%) and prediction (17%). Color fundus photography was the most common imaging modality (57%), followed by fluorescein angiography (26%), with fewer studies using optical coherence tomography or optical coherence tomography angiography. Internal validation metrics were generally high (accuracy 0.79-0.99, sensitivity 0.67-1.00, specificity 0.80-1.00), but performance declined in external validation (accuracy 0.39-0.98, sensitivity 0.38-0.93), indicating limited generalizability. Segmentation models achieved Dice coefficients between 0.82 and 0.94. Only 30% of studies used external datasets, and one performed clinical validation. Explainable AI techniques were applied in 39% of studies, mostly Grad-CAM, though often in a qualitative manner. Overall, AI systems demonstrate strong potential for assisting in RVO diagnosis and management, but challenges remain. Limited dataset diversity, lack of multimodal fusion, and minimal clinical validation restrict real-world applicability. Future research should prioritize multicenter datasets, standardized evaluation, interpretability, and ethical governance to enable safe and effective integration of AI tools in ophthalmic care.